Artificial Intelligence

In 1988, I graduated from Leiden University with a major in astronomy and a minor in artificial intelligence. A Leiden diploma is traditionally written in Latin. Disciplinam principalem: Astronomiam posed no problem, as it was used about 10 times a year. But Ceteram disciplinam: Intellegentiam Artificialem was a first, so the Faculty of Classical Languages had to be consulted. Today, you’d simply ask ChatGPT.

The term ‘artificial intelligence’ was first used in 1955 in a research proposal by four scientists, including John McCarthy and Claude Shannon. At the time, the benchmark for intelligence was, of course, the human mind, and the aim was to make computers just as intelligent in specific areas. Computers were already far better (faster) at arithmetic than humans, and it was expected that they would eventually match and surpass us in other domains as well.

Because it is difficult to judge the quality of a poem written by a computer, early research focused on intelligent games like chess, where performance is objectively measurable. After Claude Shannon published his groundbreaking Programming a Computer for Playing Chess paper in 1950, the race was on to achieve that goal. It wasn’t until 1957 that Alex Bernstein succeeded in creating a fully functioning chess programme. It could play by the rules, but it was no match for even a weak human opponent. Forty years later, Deep Blue defeated world champion Garry Kasparov in a six-game match.

As computers got better at chess, humans gradually ceased to be the benchmark for – chess – intelligence

For my minor in artificial intelligence, I developed a method to extract chess knowledge from games between strong players. At the time, the best computers played at the level of an International Master, so they still had plenty to learn from Grandmasters. The idea was to treat games between Grandmasters as ‘crystallised knowledge’ that could be mined. This resulted in a hefty thesis from which I distilled two papers, both published in the International Computer Chess Association Journal. They were my first scientific publications.

As computers got better at chess, humans gradually ceased to be the benchmark for – chess – intelligence. In endgames with few pieces, computers had been dominant since the 1980s, overturning long-held theories. The Encyclopaedia of Chess Openings was also shown to contain many inaccuracies. It became increasingly less obvious that computers had anything left to learn from humans. The ultimate milestone was achieved in 2017 when AlphaZero (developed by DeepMind, a Google subsidiary since 2014) learned to play chess by playing 44 million games against itself. The programme only knew the rules of the game and developed its playing style free of any ‘human biases’. Nobody can now defeat AlphaZero.

What artificial intelligence has demonstrated in chess is likely to happen in many – perhaps even all? – other fields where we currently believe humans are superior. Computers will match and then far surpass human intelligence. So far, that our puny brains can no longer grasp what the computer is capable of. What ChatGPT and other AI models are achieving today is already impressive, but this is just the beginning. In 1957, everyone laughed at Bernstein’s chess programme, and the question “when will a computer beat the world champion?” was considered ludicrous. Forty years on, it happened. Beati ignari.